Classification bias

Reviewing Table 15-1 after consideration of spectrum and classification bias indicates considerable error in the estimates quoted if they were to be used for annual tests on patient groups consisting largely of BPH patients. A better estimate comes from the data of Chan and colleagues from a comparison of patients with BPH and patients with known prostatic carcinoma, The number of carcinoma patients correctly predicted is still an overestimate, because this... 

The classification of objects is based on the threshold value, 0, of NET(x), also called the bias. The procedure can be described by means of a transfer function, F (Fig. 44.4a). The weighted sum of the input values of x is transmitted through a... 

In chapter 1, we discussed differences in classification strategies. The strengths and implications of the use of categorical versus dimensional methods of classification were reviewed. In the United States, there has historically been a strong bias toward dimensional typologies (Meehl, 1992). As would be expected, the DSM has been criticized by proponents of a dimensional approach (Carson, 1991). If certain forms of psychopathology are best viewed as dimensional phenomena, the current diagnostic dichotomies are arbitrary. 

The typical person is influenced by both an optimistic bias and an illusion of control (Frewer et al., 1994). When these are lost or compromised, panic can occur. While panic disorder can afflict an individual, panic also occurs as a collective phenomena. While it has been thought that emotional instability will lead some people to be more likely to panic, this is not always found. However, people who are emotionally unstable are more likely to attach importance to information provided during the crises than more emotionally stable individuals (Verbeke and Van Kenhove, 2002). Furthermore, some people are more likely to believe they are sick or affected than others. Feldman et al. (1999) examined the panic predisposition of people based on their classification by the big five personality factors. People were inoculated with a common cold virus, and those who were classified as neurotic were more likely to report unfounded illness and more symptoms than other groups. In contrast to this, openness to experience was associated with reporting unfounded symptoms in those with verifiable colds, whereas conscientiousness was associated with reporting unfounded illness in those who were not ill (Feldman et al, 1999). 

As in the above example in which the wine quality is a somewhat ill-defined concept subject to individual taste, many classification schemes are often heavily biased by the viewpoint of the researcher (and this can influence the performance). Unsupervised learning (e.g. Ritter et al., 1992) largely avoids this bias but at the cost of often less powerful methods and the missing interpretation of the arising classes, where often it is not obvious what these classes represent. 

Simon et al. (14) also showed that cross-validating the prediction rule after selection of differentially expressed genes from the full data set does little to correct the bias of the re-substitution estimator 90.2% of simulated data sets with no true relationship between expression data and class still result in zero misclassifications. When feature selection was also re-done in each cross-validated training set, however, appropriate estimates of mis-classification error were obtained the median estimated misclassification rate was approximately 50%. 

The development of the relatively youthful class of polymetallic complexes has depended on the advent of routine X-ray structure determination, and consequently the bulk of the information and understanding at present involves geometrical structure. This metrical bias, reinforced by the intriguing unpredictability of many aggregate structures, determines the content and organization of this chapter. The primary classification of compounds is structural, according to increasing numbers of metal atoms. 

Study bias may be selection bias or information bias. Selection bias may occur in the choice of subjects for the study (e.g. exclusion of individuals who are not fluent in a particular language). Selection bias may also result from an individual s reluctance to participate in a study owing to concerns over a perceived exposure, resultant health effect, or educational and socioeconomic status of the participants. Parents who perceive that an exposure in their child s environment may have resulted in an adverse health effect may feel responsible for not protecting their child. Information bias may result from inappropriate classification of the individual study participants or from the information provided. For example, interview bias may result when an interviewer is not blind to the exposure of the test population. Recall bias may result when participants with specific exposures or effects respond differently from those without the specific exposures or effects. 

If the agreement is satisfactory, the classification procedure can be used on a routine bias. If not, some of the classification algorithm parameters, the composition of the training population, the class characteristics or the selection of the shape descriptors should be modified. 

Differential misclassification occurs when the classification of disease is dependent on the exposure status or the classification of exposure is dependent on the disease status. Differential misclassification can bias the RR in either direction, and often the direction is unknown. Some examples of differential misclassification of exposure are recall bias and observer bias. Recall bias, which is limited to case-control studies, occurs when the cases remember exposure differently than healthy controls this type of bias usually results in finding a greater effect than what is real. Observer bias can occur if the observers, such as study interviewers, incorrectly assign exposure because they know the outcome status of an individual, or it can occur in the follow-up of disease if the observer knows the exposure status of the subject. Ideally, the observer should be blind to the outcome or exposure status of the study subjects. 

A multilayer perception with two hidden units is shown in Figure 3.7, with the actual weights and bias terms included, after training. This network can solve the nonlinear depression classification problem, presented in Figure 3.4. 

For human evidence to provide the primary basis for a Category lA classification there must be reliable evidence of an adverse effect on reproduction in humans. Evidence used for classification should ideally be firom well conducted epidemiological studies which include the use of appropriate controls, balanced assessment, and due consideration of bias or confounding factors. Less rigorous data from studies in humans should be supplemented with adequate data firom studies in experimental animals and classification in Category IB should be considered. 

Figure 9.3 presents one example of the verification results for the DMI-HIRLAM-U01 research model with a 1 A-km resolution for May 2005 (Mahura et al., 2005 [390]). It shows better prediction of the diurnal cycle of the average wind velocity at 10 m than with the S05-version. On average, the bias for both models was around 1.5 m/s. The verification runs underlined that increasing the resolution (down to 1 km) brings some improvement to the skill of the meteorological forecast. Nevertheless, it will be also very important for further improvements to have more detailed surface features databases and to increase the quality of the land-use classification (LUC) for urban areas. 

---